On Mon, 2014-05-12 at 21:58 +0300, Thomas Evangelidis wrote:
> Dear AMBER community,
>
> Is there any other advantage of having a GPU with big memory apart from the
> maximum system size it can endure?
>
> http://ambermd.org/gpus/#supported_gpus
>
> I am not sure if it's worth running replica exchange on a single GPU if no
> sufficient CPU resources are available.
Unless you're testing something, it's never worthwhile to run REMD on a
single GPU in my opinion.
The CPU has little to do with it. You'll see a *little* speedup (maybe)
if you have 8 CPUs and 1 GPU compared to only 1 CPU, but since the CPU
just shouts instructions to the GPU (which does _all_ of the heavy
lifting), CPU shortage is not your problem... GPU shortage is.
> For example, if I have a single
> workstation with an i7 processor (8 threads) and a GTX 680 2GB will it be
> much faster running 8 replicas of Hamiltonian Replica Exchnage on the GPU
> for a system of <100 K atoms?
Much faster than what? Running on the CPU? Maybe a bit faster, you
would have to try it. Note that i7 CPUs typically have 4 cores with
'hyperthreading', so it looks like 8 cores to the OS. It's not 8 cores,
though, and with heavy double-precision applications like Amber, the
benefits of hyperthreading are limited. I think they will both be slow,
though (as in you could probably get more sampling done with
consecutive, independent calculations).
> Will I be able to use more replicas if the
> GPU has 4 GB of memory?
Yes, but it will still be very slow. The more replicas you add, the
slower it will become (and I doubt the slowdown will be linear).
HTH,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon May 12 2014 - 12:30:09 PDT
